Self-Recognition in Language Models.

This repository contains the official implementation for the paper Self-Recognition in Language Models [1].

• 🎓 full paper available on arXiv
• 📝 we also wrote a blog post

---

Overview

In our paper we proposed assessing self-recognition in language models (LMs) using model-generated security questions. This approach takes three steps:

1. ❓ generate a set of questions;
2. 💬 generate a set of answers to these questions;
3. ⚖️ generate "verdicts" by showing LMs questions with n-answers, and prompting them to select their own.

This repository contains code to reproduce the experiments of the paper and is structured as follows:

.
├── src/
│   ├── models/ (currently support: Anthropic, Cohere, Google, Microsoft, OpenAI, TogetherAI)
│   ├── configs/ (configurations to create (i) questions, (ii) answers, (iii) verdicts)
│   └── *.py
├── data/
│   ├── api_settings/
│   ├── model_settings/
│   ├── prompts/
│   ├── questions/
│   ├── responses/
│   ├── verdicts/
│   └── llm_model_details.yaml
├── secrets.json (to be created)
└── gcp_secrets.json (optional)

A limited set of example questions, answers, and verdicts are provided in the data/ directory.

Usage Steps

We use hydra to manage configurations. The main entry point is src/run.py, which takes a configuration file as input. Configuration files are stored in src/configs/ and are used to generate questions, answers, and verdicts:

src/
└── configs/
    ├── generate_questions.yaml
    ├── generate_responses.yaml
    └── generate_verdicts.yaml

Simply navigate to these files to specify the model(s) you want to use. LM wrappers for most leading providers are included in src/models/.

After generating questions, responses, and verdicts, hydra will save the output to a specified directory, logs/ per default.

questions:

• to generate: python src/run.py defaults.experiments=generate_questions
• saves a questions.csv file to logs/<your-experiment>
• copy this file to data/questions/ for the next step

responses:

• to generate: python src/run.py defaults.experiments=generate_responses
• saves a responses.csv file to logs/<your-experiment>
• copy this file to data/responses/ for the next step

verdicts:

• to generate: python src/run.py defaults.experiments=generate_verdicts
• saves a verdicts.csv file to logs/<your-experiment>
• copy this file to data/verdicts/ for the next step

evaluation:

• to process verdicts and make sure they are correctly formatted: python src/verdict_evaluation.py --base_folder=<path-to-verdicts>
• this creates verdicts_extracted.csv in the same directory
• to evaluate the performance of the model: python src/evaluations.py --base_folder=<path-to-extacted-verdicts>

Having run these steps, you can use various tools to analyze the results. For example, see the files:

• src/analys.py
• src/visualization.py

Setup

The simplest way to get started is to:

1. clone this repository, then
2. create a secrets.json file in the root directory with the following structure:

{
    "openai": {
        "api_key": "<your-key>"
    },
    "azure": {
        "api_key": "<your-key>"
    },
    "anthropic": {
        "api_key": "<your-key>"
    },
    "google": {
        "api_key": "<your-key>"
    },
    "cohere": {
        "api_key": "<your-key>"
    },
  "together_ai": {
      "api_key": "<your-key>"
  }
}

In this file, insert your own API key for one of the following providers: {Anthropic, Cohere, Google, OpenAI, Microsoft}. This secrets.json file is part of the .gitignore, to prevent you from accidentally pushing your raw keys to GitHub :). (see 'note' below if using Google/Azure models)

Next, create a virtual environment and install the packages listed in requirements.txt. Once this is done you're all set.

For any questions, feel free to open a ticket or reach out directly to Tim :).

Note on Google/MSFT Azure

If you are using Google or MSFT Azure, you also need to update the relevant endpoints in data/api_settings/apis.yaml. At the time of release, the Google Vertex API does not support simple API keys in all regions. To get around this, you have to create a (1) service account, (2) set some permissions, (3) download a .json. Save the exported .json file in a file called gcp_secrets.json in the root directory of this project (also in .gitignore). See the following docs for a walkthrough.

License

MIT

Citation

Please cite our work using one of the following if you end up using the repository - thanks!

[1] T.R. Davidson, V. Surkov, V. Veselovsky, G. Russo, R. West, C. Gulcehre. 
Self-Recognition in Language Models. arXiv preprint, arXiv:2407.06946, 2024.

BibTeX format:

@article{davidson2024selfrecognitionlanguagemodels,
      title={Self-Recognition in Language Models}, 
      author={Tim R. Davidson and 
              Viacheslav Surkov and 
              Veniamin Veselovsky and 
              Giuseppe Russo and 
              Robert West and 
              Caglar Gulcehre},
      year={2024},
      journal={EMNLP},
      url={https://arxiv.org/abs/2407.06946}
}